Apple AI Updates: October 7, 2026

1. Apple’s RISED Uses Rubrics to Pick Training Data and Guide Self-Distillation for Agents

Apple. RISED targets LLM agents trained across many interactive environments, where scalar rewards say little about how behaviors relate across environments and give weak contrast within a rollout group. An LLM judge tags each rollout with a rubric vocabulary shared across environments, and those behavioral profiles drive batch selection to balance composition and cut redundancy; positive rubrics then serve as privileged context for an on-policy self-distillation teacher that adds token-level supervision, while negative rubrics steer generation away from recurring failure modes. Apple reports that RISED achieves the highest mean pass rate across environments on multiple model backbones and ranks first or second in every individual environment, which makes rubric tagging a practical alternative to hand-tuned environment mixing in multi-task agent RL. Source